Tensor Decomposition and Memory Networks for SAT Reading Comprehension

Abstract

Abstract We present a general approach for machine comprehension tasks by converting the text to a knowledge graph and the questions to queries on the graph. We extend [19] and use the Stanford NLP Toolkit’s Dependency Parser [17, 9] to transform each sentence into a set of entity-relation triples. We use word2vec [18] to convert the questions into queries on the graph. We present a tensor decomposition approach to answering queries by adding Semantically Smooth Embedding [11] to RESCAL [20]. We also generalize the Memory Networks [28, 25] architecture to take any knowledge graph as input. We evaluate these models on three full SAT reading comprehension tests. The models presented here outperform their respective baselines. Both models demonstrate the ability to capture the semantic and structural information in the text and answer questions using that information.